Performance testing via chron jobs

[jpmorgan98]

In the test/performance directory there is currently only one test. I am thinking we just start with whatever that test is on a given number of quartz nodes, then go from there, but I go some other questions:

• Are we going to be comparing the solutions for correctness? If so where are we storing that tally data (which might be massive)
• What are the jobs (c5g7 pulsed sphere)
• What are the job parameters we are shooting for (machine, # of nodes, # MPI ranks, Numba v Python mode, etc.)
• Where will we be storing the job runtimes (I say another GitHub repo in the CEMeNT)

[ilhamv]

I think we don't need to check for correctness, as that is covered by regression and verification tests. But we still need to run problems with huge tally sizes.

For the runtime record, I think we can store it as a text file in the repo. We will also add a Python script that generates plots of the runtime record. As I think more about it, we may want to exclude the performance test from automatic testing. Instead, we should run it manually and commit new runtime records as a PR as needed.

I'll make lists of the test problems and the performance metrics and post them here (going to refer to our "metrics of victory").

[ilhamv]

Performance test problems

SMR

• k-eigenvalue
• Four-phase transient
• Off-critical transient

C5G7

• k-eigenvalue
• Four-phase transient
• Off-critical transient

Transient shielding

• Kobayashi dog-leg
• Pulsed sphere

Burst transient

• Dragon

[jpmorgan98]

Got it! For this point:

As I think more about it, we may want to exclude the performance test from automatic testing.

We could do this as nightly or weekly jobs. This would also help burn some cycles on the machines we have access to.

Also any idea what the number of nodes/ threads we want to do for each of these?